Table 5 The performance comparison above considers four methods for computing Problem 4 in Table 2

From: Solving Boltzmann optimization problems with deep learning

Model

Time to compute 100 values

Average time to compute 1 value

SLSQP (approx grad)

4 days, 5:01:24.4

60.6 min

SLSQP (explicit grad)

5:08:53.7

3.09 min

Random forest regressor

31.987 ms after 258 s training

320 ms

DJINN DNN

28.5 ms after 251 s training

285 ms

  1. It shows the time required to calculate the min–max Boltzmann probability for an ensemble of 100 computations starting from different auxiliary spins. Half of the auxiliary spins were randomly selected from values that satisfy the constraints of Eq. (2) while the other half were selected from values that did not. The final column extrapolates the nominal average times to perform a single Boltzmann probability prediction for any randomly chosen set of initial auxiliary spins.The SLSQP solvers and random forest regressors were run in serial on an AMD EPYC 7713 64-Core processor running at 2 GHz. The DNNs were run on a single NVDIA A6000 GPU. DJINN has an additional overhead in determining an optimum number of parameters.
Back to article page